People say LLMs need too much steering, that they can’t comprehend problems. As someone who spent most of their academic pursuits on Requirements Engineering, let me tell you something: humans are also very bad at following instructions and understanding problems.

50+ years of RE research basically concern themselves with human-to-human prompt engineering.

The inherent tension of agentic coding is between delegation and comprehension. If I need to fully understand, I cannot delegate — at least not with the expectation of saving time.

The question will be: how to build agentic development environments and processes that let you delegate without losing control? They must establish trust.